I Built a Weekly SEO Review With an AI Agent. Here’s What It Can—and Can’t—Decide.

I maintain a website built with Astro. Every week, Hermes Agent reviews its search data and puts a report in a GitHub pull request. I read the recommendations, decide which ones make sense, and make the changes myself, often with Claude helping me edit the code and content.

That division of work is deliberate. The agent analyzes the data; it doesn’t decide what the website should say.

Three sources, three different questions

The review uses Google Search Console, Bing Webmaster Tools, and GA4. They’re useful together, but they don’t measure the same thing.

Google Search Console shows how pages appear in Google Search: queries, impressions, clicks, CTR, and average position. It helps me spot a page that’s getting visibility without many clicks, or a query that might deserve a closer look.

Bing Webmaster Tools provides another view of search performance and indexing. I don’t assume that a pattern in Google will also appear in Bing.

GA4 helps me look at what happens after people reach the site. I keep its metrics separate from search metrics: total GA4 sessions are not organic-search sessions, and neither number explains on its own why someone visited.

The agent accesses these sources through GSC Magic. The point isn’t to ask an LLM to “do SEO.” It’s to let it inspect the underlying data before recommending what I should investigate.

Why the report lives in a pull request

A chat response is easy to read once and lose. A GitHub PR gives the report a reviewable place alongside the Astro site.

I can inspect the reasoning, question a recommendation, and decide what belongs on the site. The PR is a handoff, not an automatic publishing step. After reading it, I might change a page, investigate a technical issue, or do nothing.

The most useful recommendations connect a specific page or query to a specific signal. “Improve this page’s SEO” gives me little to work with. “Clicks fell while impressions stayed relatively steady; inspect the queries and the current search snippet” gives me a place to start. It still isn’t a diagnosis I should accept without checking.

Data can suggest a change. It can’t approve one.

An agent can identify a query with substantial impressions. It can’t decide whether that query represents the audience I want to reach.

Writing for the wrong query might increase visits while making the site less useful to prospective customers. Conversely, a page with modest traffic may answer a question that matters deeply to the right visitor.

I also check every proposed title or content change against what the product actually does. A more clickable title isn’t an improvement if it promises something the page cannot deliver.

That’s why I review the report myself and make site changes separately. Claude may help me implement an approved edit, but the recommendation doesn’t become a change just because the agent produced it.

A weekly report is not proof of impact

Search data is noisy. For a smaller site, a difference of a few clicks can look dramatic as a percentage. Search Console data can arrive with a delay. Average position and CTR can shift because the mix of queries changed, even if the page didn’t.

GA4 adds context, but it doesn’t solve attribution. A rise in total sessions might have nothing to do with organic search. Even if organic visits increase after I edit a page, that sequence alone doesn’t prove the edit caused the increase. Competitors, search results, seasonality, and other pages can all change at the same time.

I treat the weekly PR as a list of investigations, not a scorecard proving the agent was right. If the evidence is thin, waiting for more data can be a better decision than rewriting a page.

The boundary that makes this useful

The workflow works for me because each step has a clear owner:

1. Search and analytics tools provide the evidence.
2. Hermes Agent reviews it and presents recommendations in a GitHub PR.
3. I assess the recommendations against the product, the page, and the strength of the evidence.
4. I make the changes I approve.

It doesn’t make SEO automatic, and I can’t attribute a traffic gain to the agent without stronger evidence. What it gives me is a repeatable way to inspect the data and notice questions worth investigating—without handing editorial judgment to a model.